Robust and Low-Rank Representation for Fast Face Identification With Occlusions

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Face identification systems often struggle with occlusions, particularly block occlusions that obscure significant portions of facial images.
• Existing robust representation methods are primarily designed for non-contiguous errors like pixel corruption, limiting their effectiveness against contiguous block occlusions.

Purpose of the Study:

• To propose a novel iterative method for robust face identification in the presence of block occlusions.
• To develop a representation that effectively models contiguous errors with spatial continuity.

Main Methods:

• An iterative approach utilizing a robust representation characterized by a tailored loss function for error distribution and a low-rank structure for error images.
• Employing the alternating direction method of multipliers (ADMM) for computational efficiency.

Main Results:

• The proposed joint characterization effectively models errors with spatial continuity, outperforming existing methods in face identification.
• Demonstrated significant improvements in both identification rates and computational time on diverse face databases.

Conclusions:

• The developed iterative method offers a computationally efficient and effective solution for face identification under block occlusion.
• This approach advances robust representation techniques for handling contiguous errors in biometric systems.